Microsoft thinks a face makes someone friendly. This fall, the company quietly began testing “Portraits” — a set of animated, stylized human avatars that talk, move their heads, and make natural facial expressions while you carry on a voice conversation with Copilot. The experiment lives in Copilot Labs and is currently rolling out only to users in the United States, the United Kingdom and Canada.

Why a face?

If you’ve ever nervously talked to a phone tree or politely said “thanks” to a voice assistant, you already know voice alone doesn’t always feel like a conversation. Microsoft AI chief Mustafa Suleyman framed the new test exactly that way: some people told the team “they’d feel more comfortable talking to a face when using voice,” so Portraits exists to see whether an animated face actually lowers the friction in a real exchange.

That’s the modest pitch: let Copilot show up as a familiar visual cue so spoken back-and-forth feels less like dictation and more like dialogue.

What Portraits actually is

Portraits gives users about 40 stylized human avatars to choose from, and you can pair each portrait with different voices. During a live voice session, the portrait generates synchronized lip movement, head turns and micro-expressions in real time, so the image looks like it’s “listening” and responding — not just lip-syncing to audio. Microsoft says the portraits are intentionally non-photorealistic (think stylized, not uncanny) and that the underlying Copilot intelligence and safety filters remain the same.

Availability is limited. Microsoft has restricted the feature to a subset of Copilot users aged 18 and over and placed daily and session-based time caps on usage; the company also requires clear on-screen indicators that you’re talking to an AI, not a human. Those guardrails suggest Microsoft is trying to collect honest user behavior data while keeping the scope tightly controlled.

The tech behind the eyes

The lifelike motion in these portraits is powered by a technology Microsoft Research calls VASA-1, a model trained to turn a single static image and an audio stream into a talking, expressive face — in real time. VASA’s creators describe it as a system that goes beyond just matching lips to sound; it models natural head motion and a range of facial “affective” signals that make an avatar feel alive. Because it works from a single picture, there’s no need for complex 3-D modeling — which makes it fast and more practical for live chat.

That single-image capability is a double-edged sword: it’s what makes the feature easy to deploy, but it’s also the part that raises predictable concerns about deepfakes and misuse. Microsoft’s response so far has been to keep Portraits stylized and gated, while continuing to emphasize safety tools baked into Copilot.

Where this fits in Microsoft’s playbook

Portraits isn’t the first time Microsoft has experimented with giving Copilot a face. Earlier this year, the company tested more cartoonish options under the “Copilot Appearances” umbrella — whimsical, fantastical looks meant to test user preference for non-human visuals. Portraits feels like a complementary experiment: same idea (visual persona + voice), different aesthetic (human-leaning vs. cartoon). The experiments are being run inside Copilot Labs and Copilot Studio prototypes as Microsoft tries to learn which cues make people engage — and which make them uncomfortable.

There are product incentives, too: several reports note these Labs features are aimed at Copilot Pro users and premium subscribers — roughly the same playbook used to monetize more advanced, “human-like” interactions.

The safety shadow: why Microsoft is cautious

Microsoft’s incremental rollout and strict age and time limits read like a lesson learned from the industry’s messy experiments with personified AIs. Rivals such as xAI’s Grok have already pushed companion-style avatars and “spicy” modes that flirt dangerously close to sexualized or exploitative interactions — features that drew scrutiny from journalists, moderators and internal teams. Meanwhile, services like Character .AI have faced lawsuits and investigations over harmful interactions and the use of copyrighted characters. Those controversies have made platforms skittish about releasing any avatar-based social AI without substantial safety nets.

Microsoft’s answer has been a mix of product design (stylized, not photorealistic), guardrails (age caps, time limits, clear AI labeling) and an emphasis on safety features already built into Copilot. Whether that will be enough — especially if people try to push the avatars into flirtatious or manipulative uses — is an open question the company clearly intends to study before wider release.

Why this matters

We’re watching a small evolution in how AI products present themselves. If you strip away the marketing, Portraits is an experiment in social affordance: can a little visual humanity — carefully designed and tightly policed — make automated systems easier to use without causing new harms? Microsoft is betting yes, but is proceeding slowly because the lesson from the last few years is that personified AI can quickly generate messy, emotional, and sometimes dangerous interactions when left unchecked.

If the experiment scales, expect to see more nuanced choices about identity, expression and control — from user-customizable avatars to corporate policies about what is allowed in “companion” modes. For now, Portraits is a measured step: a face, not a personality — and a test to see whether people actually prefer one.